xref: /linux/arch/mips/mm/tlbex.c (revision fba2689ee77e63b05e203b3f26079ef915e55660)
1 /*
2  * This file is subject to the terms and conditions of the GNU General Public
3  * License.  See the file "COPYING" in the main directory of this archive
4  * for more details.
5  *
6  * Synthesize TLB refill handlers at runtime.
7  *
8  * Copyright (C) 2004, 2005, 2006, 2008	 Thiemo Seufer
9  * Copyright (C) 2005, 2007, 2008, 2009	 Maciej W. Rozycki
10  * Copyright (C) 2006  Ralf Baechle (ralf@linux-mips.org)
11  * Copyright (C) 2008, 2009 Cavium Networks, Inc.
12  * Copyright (C) 2011  MIPS Technologies, Inc.
13  *
14  * ... and the days got worse and worse and now you see
15  * I've gone completely out of my mind.
16  *
17  * They're coming to take me a away haha
18  * they're coming to take me a away hoho hihi haha
19  * to the funny farm where code is beautiful all the time ...
20  *
21  * (Condolences to Napoleon XIV)
22  */
23 
24 #include <linux/bug.h>
25 #include <linux/export.h>
26 #include <linux/kernel.h>
27 #include <linux/types.h>
28 #include <linux/smp.h>
29 #include <linux/string.h>
30 #include <linux/cache.h>
31 #include <linux/pgtable.h>
32 
33 #include <asm/cacheflush.h>
34 #include <asm/cpu-type.h>
35 #include <asm/mmu_context.h>
36 #include <asm/uasm.h>
37 #include <asm/setup.h>
38 #include <asm/tlbex.h>
39 
40 static int mips_xpa_disabled;
41 
42 static int __init xpa_disable(char *s)
43 {
44 	mips_xpa_disabled = 1;
45 
46 	return 1;
47 }
48 
49 __setup("noxpa", xpa_disable);
50 
51 /*
52  * TLB load/store/modify handlers.
53  *
54  * Only the fastpath gets synthesized at runtime, the slowpath for
55  * do_page_fault remains normal asm.
56  */
57 extern void tlb_do_page_fault_0(void);
58 extern void tlb_do_page_fault_1(void);
59 
60 struct work_registers {
61 	int r1;
62 	int r2;
63 	int r3;
64 };
65 
66 struct tlb_reg_save {
67 	unsigned long a;
68 	unsigned long b;
69 } ____cacheline_aligned_in_smp;
70 
71 static struct tlb_reg_save handler_reg_save[NR_CPUS];
72 
73 static inline int r45k_bvahwbug(void)
74 {
75 	/* XXX: We should probe for the presence of this bug, but we don't. */
76 	return 0;
77 }
78 
79 static inline int r4k_250MHZhwbug(void)
80 {
81 	/* XXX: We should probe for the presence of this bug, but we don't. */
82 	return 0;
83 }
84 
85 extern int sb1250_m3_workaround_needed(void);
86 
87 static inline int __maybe_unused bcm1250_m3_war(void)
88 {
89 	if (IS_ENABLED(CONFIG_SB1_PASS_2_WORKAROUNDS))
90 		return sb1250_m3_workaround_needed();
91 	return 0;
92 }
93 
94 static inline int __maybe_unused r10000_llsc_war(void)
95 {
96 	return IS_ENABLED(CONFIG_WAR_R10000_LLSC);
97 }
98 
99 static int use_bbit_insns(void)
100 {
101 	switch (current_cpu_type()) {
102 	case CPU_CAVIUM_OCTEON:
103 	case CPU_CAVIUM_OCTEON_PLUS:
104 	case CPU_CAVIUM_OCTEON2:
105 	case CPU_CAVIUM_OCTEON3:
106 		return 1;
107 	default:
108 		return 0;
109 	}
110 }
111 
112 static int use_lwx_insns(void)
113 {
114 	switch (current_cpu_type()) {
115 	case CPU_CAVIUM_OCTEON2:
116 	case CPU_CAVIUM_OCTEON3:
117 		return 1;
118 	default:
119 		return 0;
120 	}
121 }
122 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
123     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
124 static bool scratchpad_available(void)
125 {
126 	return true;
127 }
128 static int scratchpad_offset(int i)
129 {
130 	/*
131 	 * CVMSEG starts at address -32768 and extends for
132 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
133 	 */
134 	i += 1; /* Kernel use starts at the top and works down. */
135 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
136 }
137 #else
138 static bool scratchpad_available(void)
139 {
140 	return false;
141 }
142 static int scratchpad_offset(int i)
143 {
144 	BUG();
145 	/* Really unreachable, but evidently some GCC want this. */
146 	return 0;
147 }
148 #endif
149 /*
150  * Found by experiment: At least some revisions of the 4kc throw under
151  * some circumstances a machine check exception, triggered by invalid
152  * values in the index register.  Delaying the tlbp instruction until
153  * after the next branch,  plus adding an additional nop in front of
154  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
155  * why; it's not an issue caused by the core RTL.
156  *
157  */
158 static int m4kc_tlbp_war(void)
159 {
160 	return current_cpu_type() == CPU_4KC;
161 }
162 
163 /* Handle labels (which must be positive integers). */
164 enum label_id {
165 	label_second_part = 1,
166 	label_leave,
167 	label_vmalloc,
168 	label_vmalloc_done,
169 	label_tlbw_hazard_0,
170 	label_split = label_tlbw_hazard_0 + 8,
171 	label_tlbl_goaround1,
172 	label_tlbl_goaround2,
173 	label_nopage_tlbl,
174 	label_nopage_tlbs,
175 	label_nopage_tlbm,
176 	label_smp_pgtable_change,
177 	label_r3000_write_probe_fail,
178 	label_large_segbits_fault,
179 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
180 	label_tlb_huge_update,
181 #endif
182 };
183 
184 UASM_L_LA(_second_part)
185 UASM_L_LA(_leave)
186 UASM_L_LA(_vmalloc)
187 UASM_L_LA(_vmalloc_done)
188 /* _tlbw_hazard_x is handled differently.  */
189 UASM_L_LA(_split)
190 UASM_L_LA(_tlbl_goaround1)
191 UASM_L_LA(_tlbl_goaround2)
192 UASM_L_LA(_nopage_tlbl)
193 UASM_L_LA(_nopage_tlbs)
194 UASM_L_LA(_nopage_tlbm)
195 UASM_L_LA(_smp_pgtable_change)
196 UASM_L_LA(_r3000_write_probe_fail)
197 UASM_L_LA(_large_segbits_fault)
198 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
199 UASM_L_LA(_tlb_huge_update)
200 #endif
201 
202 static int hazard_instance;
203 
204 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
205 {
206 	switch (instance) {
207 	case 0 ... 7:
208 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
209 		return;
210 	default:
211 		BUG();
212 	}
213 }
214 
215 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
216 {
217 	switch (instance) {
218 	case 0 ... 7:
219 		uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
220 		break;
221 	default:
222 		BUG();
223 	}
224 }
225 
226 /*
227  * pgtable bits are assigned dynamically depending on processor feature
228  * and statically based on kernel configuration.  This spits out the actual
229  * values the kernel is using.	Required to make sense from disassembled
230  * TLB exception handlers.
231  */
232 static void output_pgtable_bits_defines(void)
233 {
234 #define pr_define(fmt, ...)					\
235 	pr_debug("#define " fmt, ##__VA_ARGS__)
236 
237 	pr_debug("#include <asm/asm.h>\n");
238 	pr_debug("#include <asm/regdef.h>\n");
239 	pr_debug("\n");
240 
241 	pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
242 	pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
243 	pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
244 	pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
245 	pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
246 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
247 	pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
248 #endif
249 #ifdef _PAGE_NO_EXEC_SHIFT
250 	if (cpu_has_rixi)
251 		pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
252 #endif
253 	pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
254 	pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
255 	pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
256 	pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
257 	pr_debug("\n");
258 }
259 
260 static inline void dump_handler(const char *symbol, const void *start, const void *end)
261 {
262 	unsigned int count = (end - start) / sizeof(u32);
263 	const u32 *handler = start;
264 	int i;
265 
266 	pr_debug("LEAF(%s)\n", symbol);
267 
268 	pr_debug("\t.set push\n");
269 	pr_debug("\t.set noreorder\n");
270 
271 	for (i = 0; i < count; i++)
272 		pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
273 
274 	pr_debug("\t.set\tpop\n");
275 
276 	pr_debug("\tEND(%s)\n", symbol);
277 }
278 
279 /* The only general purpose registers allowed in TLB handlers. */
280 #define K0		26
281 #define K1		27
282 
283 /* Some CP0 registers */
284 #define C0_INDEX	0, 0
285 #define C0_ENTRYLO0	2, 0
286 #define C0_TCBIND	2, 2
287 #define C0_ENTRYLO1	3, 0
288 #define C0_CONTEXT	4, 0
289 #define C0_PAGEMASK	5, 0
290 #define C0_PWBASE	5, 5
291 #define C0_PWFIELD	5, 6
292 #define C0_PWSIZE	5, 7
293 #define C0_PWCTL	6, 6
294 #define C0_BADVADDR	8, 0
295 #define C0_PGD		9, 7
296 #define C0_ENTRYHI	10, 0
297 #define C0_EPC		14, 0
298 #define C0_XCONTEXT	20, 0
299 
300 #ifdef CONFIG_64BIT
301 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
302 #else
303 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
304 #endif
305 
306 /* The worst case length of the handler is around 18 instructions for
307  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
308  * Maximum space available is 32 instructions for R3000 and 64
309  * instructions for R4000.
310  *
311  * We deliberately chose a buffer size of 128, so we won't scribble
312  * over anything important on overflow before we panic.
313  */
314 static u32 tlb_handler[128];
315 
316 /* simply assume worst case size for labels and relocs */
317 static struct uasm_label labels[128];
318 static struct uasm_reloc relocs[128];
319 
320 static int check_for_high_segbits;
321 static bool fill_includes_sw_bits;
322 
323 static unsigned int kscratch_used_mask;
324 
325 static inline int __maybe_unused c0_kscratch(void)
326 {
327 	return 31;
328 }
329 
330 static int allocate_kscratch(void)
331 {
332 	int r;
333 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
334 
335 	r = ffs(a);
336 
337 	if (r == 0)
338 		return -1;
339 
340 	r--; /* make it zero based */
341 
342 	kscratch_used_mask |= (1 << r);
343 
344 	return r;
345 }
346 
347 static int scratch_reg;
348 int pgd_reg;
349 EXPORT_SYMBOL_GPL(pgd_reg);
350 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
351 
352 static struct work_registers build_get_work_registers(u32 **p)
353 {
354 	struct work_registers r;
355 
356 	if (scratch_reg >= 0) {
357 		/* Save in CPU local C0_KScratch? */
358 		UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
359 		r.r1 = K0;
360 		r.r2 = K1;
361 		r.r3 = 1;
362 		return r;
363 	}
364 
365 	if (num_possible_cpus() > 1) {
366 		/* Get smp_processor_id */
367 		UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
368 		UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
369 
370 		/* handler_reg_save index in K0 */
371 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
372 
373 		UASM_i_LA(p, K1, (long)&handler_reg_save);
374 		UASM_i_ADDU(p, K0, K0, K1);
375 	} else {
376 		UASM_i_LA(p, K0, (long)&handler_reg_save);
377 	}
378 	/* K0 now points to save area, save $1 and $2  */
379 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
380 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
381 
382 	r.r1 = K1;
383 	r.r2 = 1;
384 	r.r3 = 2;
385 	return r;
386 }
387 
388 static void build_restore_work_registers(u32 **p)
389 {
390 	if (scratch_reg >= 0) {
391 		uasm_i_ehb(p);
392 		UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
393 		return;
394 	}
395 	/* K0 already points to save area, restore $1 and $2  */
396 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
397 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
398 }
399 
400 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
401 
402 /*
403  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
404  * we cannot do r3000 under these circumstances.
405  *
406  * The R3000 TLB handler is simple.
407  */
408 static void build_r3000_tlb_refill_handler(void)
409 {
410 	long pgdc = (long)pgd_current;
411 	u32 *p;
412 
413 	memset(tlb_handler, 0, sizeof(tlb_handler));
414 	p = tlb_handler;
415 
416 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
417 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
418 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
419 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
420 	uasm_i_sll(&p, K0, K0, 2);
421 	uasm_i_addu(&p, K1, K1, K0);
422 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
423 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
424 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
425 	uasm_i_addu(&p, K1, K1, K0);
426 	uasm_i_lw(&p, K0, 0, K1);
427 	uasm_i_nop(&p); /* load delay */
428 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
429 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
430 	uasm_i_tlbwr(&p); /* cp0 delay */
431 	uasm_i_jr(&p, K1);
432 	uasm_i_rfe(&p); /* branch delay */
433 
434 	if (p > tlb_handler + 32)
435 		panic("TLB refill handler space exceeded");
436 
437 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
438 		 (unsigned int)(p - tlb_handler));
439 
440 	memcpy((void *)ebase, tlb_handler, 0x80);
441 	local_flush_icache_range(ebase, ebase + 0x80);
442 	dump_handler("r3000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x80));
443 }
444 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
445 
446 /*
447  * The R4000 TLB handler is much more complicated. We have two
448  * consecutive handler areas with 32 instructions space each.
449  * Since they aren't used at the same time, we can overflow in the
450  * other one.To keep things simple, we first assume linear space,
451  * then we relocate it to the final handler layout as needed.
452  */
453 static u32 final_handler[64];
454 
455 /*
456  * Hazards
457  *
458  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
459  * 2. A timing hazard exists for the TLBP instruction.
460  *
461  *	stalling_instruction
462  *	TLBP
463  *
464  * The JTLB is being read for the TLBP throughout the stall generated by the
465  * previous instruction. This is not really correct as the stalling instruction
466  * can modify the address used to access the JTLB.  The failure symptom is that
467  * the TLBP instruction will use an address created for the stalling instruction
468  * and not the address held in C0_ENHI and thus report the wrong results.
469  *
470  * The software work-around is to not allow the instruction preceding the TLBP
471  * to stall - make it an NOP or some other instruction guaranteed not to stall.
472  *
473  * Errata 2 will not be fixed.	This errata is also on the R5000.
474  *
475  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
476  */
477 static void __maybe_unused build_tlb_probe_entry(u32 **p)
478 {
479 	switch (current_cpu_type()) {
480 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
481 	case CPU_R4600:
482 	case CPU_R4700:
483 	case CPU_R5000:
484 	case CPU_NEVADA:
485 		uasm_i_nop(p);
486 		uasm_i_tlbp(p);
487 		break;
488 
489 	default:
490 		uasm_i_tlbp(p);
491 		break;
492 	}
493 }
494 
495 void build_tlb_write_entry(u32 **p, struct uasm_label **l,
496 			   struct uasm_reloc **r,
497 			   enum tlb_write_entry wmode)
498 {
499 	void(*tlbw)(u32 **) = NULL;
500 
501 	switch (wmode) {
502 	case tlb_random: tlbw = uasm_i_tlbwr; break;
503 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
504 	}
505 
506 	if (cpu_has_mips_r2_r6) {
507 		if (cpu_has_mips_r2_exec_hazard)
508 			uasm_i_ehb(p);
509 		tlbw(p);
510 		return;
511 	}
512 
513 	switch (current_cpu_type()) {
514 	case CPU_R4000PC:
515 	case CPU_R4000SC:
516 	case CPU_R4000MC:
517 	case CPU_R4400PC:
518 	case CPU_R4400SC:
519 	case CPU_R4400MC:
520 		/*
521 		 * This branch uses up a mtc0 hazard nop slot and saves
522 		 * two nops after the tlbw instruction.
523 		 */
524 		uasm_bgezl_hazard(p, r, hazard_instance);
525 		tlbw(p);
526 		uasm_bgezl_label(l, p, hazard_instance);
527 		hazard_instance++;
528 		uasm_i_nop(p);
529 		break;
530 
531 	case CPU_R4600:
532 	case CPU_R4700:
533 		uasm_i_nop(p);
534 		tlbw(p);
535 		uasm_i_nop(p);
536 		break;
537 
538 	case CPU_R5000:
539 	case CPU_NEVADA:
540 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
541 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
542 		tlbw(p);
543 		break;
544 
545 	case CPU_R4300:
546 	case CPU_5KC:
547 	case CPU_TX49XX:
548 	case CPU_PR4450:
549 		uasm_i_nop(p);
550 		tlbw(p);
551 		break;
552 
553 	case CPU_R10000:
554 	case CPU_R12000:
555 	case CPU_R14000:
556 	case CPU_R16000:
557 	case CPU_4KC:
558 	case CPU_4KEC:
559 	case CPU_M14KC:
560 	case CPU_M14KEC:
561 	case CPU_SB1:
562 	case CPU_SB1A:
563 	case CPU_4KSC:
564 	case CPU_20KC:
565 	case CPU_25KF:
566 	case CPU_BMIPS32:
567 	case CPU_BMIPS3300:
568 	case CPU_BMIPS4350:
569 	case CPU_BMIPS4380:
570 	case CPU_BMIPS5000:
571 	case CPU_LOONGSON2EF:
572 	case CPU_LOONGSON64:
573 	case CPU_R5500:
574 		if (m4kc_tlbp_war())
575 			uasm_i_nop(p);
576 		fallthrough;
577 	case CPU_ALCHEMY:
578 		tlbw(p);
579 		break;
580 
581 	case CPU_RM7000:
582 		uasm_i_nop(p);
583 		uasm_i_nop(p);
584 		uasm_i_nop(p);
585 		uasm_i_nop(p);
586 		tlbw(p);
587 		break;
588 
589 	case CPU_VR4111:
590 	case CPU_VR4121:
591 	case CPU_VR4122:
592 	case CPU_VR4181:
593 	case CPU_VR4181A:
594 		uasm_i_nop(p);
595 		uasm_i_nop(p);
596 		tlbw(p);
597 		uasm_i_nop(p);
598 		uasm_i_nop(p);
599 		break;
600 
601 	case CPU_VR4131:
602 	case CPU_VR4133:
603 		uasm_i_nop(p);
604 		uasm_i_nop(p);
605 		tlbw(p);
606 		break;
607 
608 	case CPU_XBURST:
609 		tlbw(p);
610 		uasm_i_nop(p);
611 		break;
612 
613 	default:
614 		panic("No TLB refill handler yet (CPU type: %d)",
615 		      current_cpu_type());
616 		break;
617 	}
618 }
619 EXPORT_SYMBOL_GPL(build_tlb_write_entry);
620 
621 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
622 							unsigned int reg)
623 {
624 	if (_PAGE_GLOBAL_SHIFT == 0) {
625 		/* pte_t is already in EntryLo format */
626 		return;
627 	}
628 
629 	if (cpu_has_rixi && !!_PAGE_NO_EXEC) {
630 		if (fill_includes_sw_bits) {
631 			UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
632 		} else {
633 			UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
634 			UASM_i_ROTR(p, reg, reg,
635 				    ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
636 		}
637 	} else {
638 #ifdef CONFIG_PHYS_ADDR_T_64BIT
639 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
640 #else
641 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
642 #endif
643 	}
644 }
645 
646 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
647 
648 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
649 				   unsigned int tmp, enum label_id lid,
650 				   int restore_scratch)
651 {
652 	if (restore_scratch) {
653 		/*
654 		 * Ensure the MFC0 below observes the value written to the
655 		 * KScratch register by the prior MTC0.
656 		 */
657 		if (scratch_reg >= 0)
658 			uasm_i_ehb(p);
659 
660 		/* Reset default page size */
661 		if (PM_DEFAULT_MASK >> 16) {
662 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
663 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
664 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
665 			uasm_il_b(p, r, lid);
666 		} else if (PM_DEFAULT_MASK) {
667 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
668 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
669 			uasm_il_b(p, r, lid);
670 		} else {
671 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
672 			uasm_il_b(p, r, lid);
673 		}
674 		if (scratch_reg >= 0)
675 			UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
676 		else
677 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
678 	} else {
679 		/* Reset default page size */
680 		if (PM_DEFAULT_MASK >> 16) {
681 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
682 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
683 			uasm_il_b(p, r, lid);
684 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
685 		} else if (PM_DEFAULT_MASK) {
686 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
687 			uasm_il_b(p, r, lid);
688 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
689 		} else {
690 			uasm_il_b(p, r, lid);
691 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
692 		}
693 	}
694 }
695 
696 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
697 				       struct uasm_reloc **r,
698 				       unsigned int tmp,
699 				       enum tlb_write_entry wmode,
700 				       int restore_scratch)
701 {
702 	/* Set huge page tlb entry size */
703 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
704 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
705 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
706 
707 	build_tlb_write_entry(p, l, r, wmode);
708 
709 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
710 }
711 
712 /*
713  * Check if Huge PTE is present, if so then jump to LABEL.
714  */
715 static void
716 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
717 		  unsigned int pmd, int lid)
718 {
719 	UASM_i_LW(p, tmp, 0, pmd);
720 	if (use_bbit_insns()) {
721 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
722 	} else {
723 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
724 		uasm_il_bnez(p, r, tmp, lid);
725 	}
726 }
727 
728 static void build_huge_update_entries(u32 **p, unsigned int pte,
729 				      unsigned int tmp)
730 {
731 	int small_sequence;
732 
733 	/*
734 	 * A huge PTE describes an area the size of the
735 	 * configured huge page size. This is twice the
736 	 * of the large TLB entry size we intend to use.
737 	 * A TLB entry half the size of the configured
738 	 * huge page size is configured into entrylo0
739 	 * and entrylo1 to cover the contiguous huge PTE
740 	 * address space.
741 	 */
742 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
743 
744 	/* We can clobber tmp.	It isn't used after this.*/
745 	if (!small_sequence)
746 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
747 
748 	build_convert_pte_to_entrylo(p, pte);
749 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
750 	/* convert to entrylo1 */
751 	if (small_sequence)
752 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
753 	else
754 		UASM_i_ADDU(p, pte, pte, tmp);
755 
756 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
757 }
758 
759 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
760 				    struct uasm_label **l,
761 				    unsigned int pte,
762 				    unsigned int ptr,
763 				    unsigned int flush)
764 {
765 #ifdef CONFIG_SMP
766 	UASM_i_SC(p, pte, 0, ptr);
767 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
768 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
769 #else
770 	UASM_i_SW(p, pte, 0, ptr);
771 #endif
772 	if (cpu_has_ftlb && flush) {
773 		BUG_ON(!cpu_has_tlbinv);
774 
775 		UASM_i_MFC0(p, ptr, C0_ENTRYHI);
776 		uasm_i_ori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
777 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
778 		build_tlb_write_entry(p, l, r, tlb_indexed);
779 
780 		uasm_i_xori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
781 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
782 		build_huge_update_entries(p, pte, ptr);
783 		build_huge_tlb_write_entry(p, l, r, pte, tlb_random, 0);
784 
785 		return;
786 	}
787 
788 	build_huge_update_entries(p, pte, ptr);
789 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
790 }
791 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
792 
793 #ifdef CONFIG_64BIT
794 /*
795  * TMP and PTR are scratch.
796  * TMP will be clobbered, PTR will hold the pmd entry.
797  */
798 void build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
799 		      unsigned int tmp, unsigned int ptr)
800 {
801 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
802 	long pgdc = (long)pgd_current;
803 #endif
804 	/*
805 	 * The vmalloc handling is not in the hotpath.
806 	 */
807 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
808 
809 	if (check_for_high_segbits) {
810 		/*
811 		 * The kernel currently implicitely assumes that the
812 		 * MIPS SEGBITS parameter for the processor is
813 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
814 		 * allocate virtual addresses outside the maximum
815 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
816 		 * that doesn't prevent user code from accessing the
817 		 * higher xuseg addresses.  Here, we make sure that
818 		 * everything but the lower xuseg addresses goes down
819 		 * the module_alloc/vmalloc path.
820 		 */
821 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
822 		uasm_il_bnez(p, r, ptr, label_vmalloc);
823 	} else {
824 		uasm_il_bltz(p, r, tmp, label_vmalloc);
825 	}
826 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
827 
828 	if (pgd_reg != -1) {
829 		/* pgd is in pgd_reg */
830 		if (cpu_has_ldpte)
831 			UASM_i_MFC0(p, ptr, C0_PWBASE);
832 		else
833 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
834 	} else {
835 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
836 		/*
837 		 * &pgd << 11 stored in CONTEXT [23..63].
838 		 */
839 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
840 
841 		/* Clear lower 23 bits of context. */
842 		uasm_i_dins(p, ptr, 0, 0, 23);
843 
844 		/* insert bit[63:59] of CAC_BASE into bit[11:6] of ptr */
845 		uasm_i_ori(p, ptr, ptr, ((u64)(CAC_BASE) >> 53));
846 		uasm_i_drotr(p, ptr, ptr, 11);
847 #elif defined(CONFIG_SMP)
848 		UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
849 		uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
850 		UASM_i_LA_mostly(p, tmp, pgdc);
851 		uasm_i_daddu(p, ptr, ptr, tmp);
852 		uasm_i_dmfc0(p, tmp, C0_BADVADDR);
853 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
854 #else
855 		UASM_i_LA_mostly(p, ptr, pgdc);
856 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
857 #endif
858 	}
859 
860 	uasm_l_vmalloc_done(l, *p);
861 
862 	/* get pgd offset in bytes */
863 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
864 
865 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
866 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
867 #ifndef __PAGETABLE_PUD_FOLDED
868 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
869 	uasm_i_ld(p, ptr, 0, ptr); /* get pud pointer */
870 	uasm_i_dsrl_safe(p, tmp, tmp, PUD_SHIFT - 3); /* get pud offset in bytes */
871 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PUD - 1) << 3);
872 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pud offset */
873 #endif
874 #ifndef __PAGETABLE_PMD_FOLDED
875 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
876 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
877 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
878 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
879 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
880 #endif
881 }
882 EXPORT_SYMBOL_GPL(build_get_pmde64);
883 
884 /*
885  * BVADDR is the faulting address, PTR is scratch.
886  * PTR will hold the pgd for vmalloc.
887  */
888 static void
889 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
890 			unsigned int bvaddr, unsigned int ptr,
891 			enum vmalloc64_mode mode)
892 {
893 	long swpd = (long)swapper_pg_dir;
894 	int single_insn_swpd;
895 	int did_vmalloc_branch = 0;
896 
897 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
898 
899 	uasm_l_vmalloc(l, *p);
900 
901 	if (mode != not_refill && check_for_high_segbits) {
902 		if (single_insn_swpd) {
903 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
904 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
905 			did_vmalloc_branch = 1;
906 			/* fall through */
907 		} else {
908 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
909 		}
910 	}
911 	if (!did_vmalloc_branch) {
912 		if (single_insn_swpd) {
913 			uasm_il_b(p, r, label_vmalloc_done);
914 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
915 		} else {
916 			UASM_i_LA_mostly(p, ptr, swpd);
917 			uasm_il_b(p, r, label_vmalloc_done);
918 			if (uasm_in_compat_space_p(swpd))
919 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
920 			else
921 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
922 		}
923 	}
924 	if (mode != not_refill && check_for_high_segbits) {
925 		uasm_l_large_segbits_fault(l, *p);
926 
927 		if (mode == refill_scratch && scratch_reg >= 0)
928 			uasm_i_ehb(p);
929 
930 		/*
931 		 * We get here if we are an xsseg address, or if we are
932 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
933 		 *
934 		 * Ignoring xsseg (assume disabled so would generate
935 		 * (address errors?), the only remaining possibility
936 		 * is the upper xuseg addresses.  On processors with
937 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
938 		 * addresses would have taken an address error. We try
939 		 * to mimic that here by taking a load/istream page
940 		 * fault.
941 		 */
942 		if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
943 			uasm_i_sync(p, 0);
944 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
945 		uasm_i_jr(p, ptr);
946 
947 		if (mode == refill_scratch) {
948 			if (scratch_reg >= 0)
949 				UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
950 			else
951 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
952 		} else {
953 			uasm_i_nop(p);
954 		}
955 	}
956 }
957 
958 #else /* !CONFIG_64BIT */
959 
960 /*
961  * TMP and PTR are scratch.
962  * TMP will be clobbered, PTR will hold the pgd entry.
963  */
964 void build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
965 {
966 	if (pgd_reg != -1) {
967 		/* pgd is in pgd_reg */
968 		uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
969 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
970 	} else {
971 		long pgdc = (long)pgd_current;
972 
973 		/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
974 #ifdef CONFIG_SMP
975 		uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
976 		UASM_i_LA_mostly(p, tmp, pgdc);
977 		uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
978 		uasm_i_addu(p, ptr, tmp, ptr);
979 #else
980 		UASM_i_LA_mostly(p, ptr, pgdc);
981 #endif
982 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
983 		uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
984 	}
985 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
986 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
987 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
988 }
989 EXPORT_SYMBOL_GPL(build_get_pgde32);
990 
991 #endif /* !CONFIG_64BIT */
992 
993 static void build_adjust_context(u32 **p, unsigned int ctx)
994 {
995 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
996 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
997 
998 	switch (current_cpu_type()) {
999 	case CPU_VR41XX:
1000 	case CPU_VR4111:
1001 	case CPU_VR4121:
1002 	case CPU_VR4122:
1003 	case CPU_VR4131:
1004 	case CPU_VR4181:
1005 	case CPU_VR4181A:
1006 	case CPU_VR4133:
1007 		shift += 2;
1008 		break;
1009 
1010 	default:
1011 		break;
1012 	}
1013 
1014 	if (shift)
1015 		UASM_i_SRL(p, ctx, ctx, shift);
1016 	uasm_i_andi(p, ctx, ctx, mask);
1017 }
1018 
1019 void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1020 {
1021 	/*
1022 	 * Bug workaround for the Nevada. It seems as if under certain
1023 	 * circumstances the move from cp0_context might produce a
1024 	 * bogus result when the mfc0 instruction and its consumer are
1025 	 * in a different cacheline or a load instruction, probably any
1026 	 * memory reference, is between them.
1027 	 */
1028 	switch (current_cpu_type()) {
1029 	case CPU_NEVADA:
1030 		UASM_i_LW(p, ptr, 0, ptr);
1031 		GET_CONTEXT(p, tmp); /* get context reg */
1032 		break;
1033 
1034 	default:
1035 		GET_CONTEXT(p, tmp); /* get context reg */
1036 		UASM_i_LW(p, ptr, 0, ptr);
1037 		break;
1038 	}
1039 
1040 	build_adjust_context(p, tmp);
1041 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1042 }
1043 EXPORT_SYMBOL_GPL(build_get_ptep);
1044 
1045 void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1046 {
1047 	int pte_off_even = 0;
1048 	int pte_off_odd = sizeof(pte_t);
1049 
1050 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1051 	/* The low 32 bits of EntryLo is stored in pte_high */
1052 	pte_off_even += offsetof(pte_t, pte_high);
1053 	pte_off_odd += offsetof(pte_t, pte_high);
1054 #endif
1055 
1056 	if (IS_ENABLED(CONFIG_XPA)) {
1057 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1058 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1059 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1060 
1061 		if (cpu_has_xpa && !mips_xpa_disabled) {
1062 			uasm_i_lw(p, tmp, 0, ptep);
1063 			uasm_i_ext(p, tmp, tmp, 0, 24);
1064 			uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1065 		}
1066 
1067 		uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1068 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1069 		UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1070 
1071 		if (cpu_has_xpa && !mips_xpa_disabled) {
1072 			uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1073 			uasm_i_ext(p, tmp, tmp, 0, 24);
1074 			uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1075 		}
1076 		return;
1077 	}
1078 
1079 	UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1080 	UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1081 	if (r45k_bvahwbug())
1082 		build_tlb_probe_entry(p);
1083 	build_convert_pte_to_entrylo(p, tmp);
1084 	if (r4k_250MHZhwbug())
1085 		UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1086 	UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1087 	build_convert_pte_to_entrylo(p, ptep);
1088 	if (r45k_bvahwbug())
1089 		uasm_i_mfc0(p, tmp, C0_INDEX);
1090 	if (r4k_250MHZhwbug())
1091 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1092 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1093 }
1094 EXPORT_SYMBOL_GPL(build_update_entries);
1095 
1096 struct mips_huge_tlb_info {
1097 	int huge_pte;
1098 	int restore_scratch;
1099 	bool need_reload_pte;
1100 };
1101 
1102 static struct mips_huge_tlb_info
1103 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1104 			       struct uasm_reloc **r, unsigned int tmp,
1105 			       unsigned int ptr, int c0_scratch_reg)
1106 {
1107 	struct mips_huge_tlb_info rv;
1108 	unsigned int even, odd;
1109 	int vmalloc_branch_delay_filled = 0;
1110 	const int scratch = 1; /* Our extra working register */
1111 
1112 	rv.huge_pte = scratch;
1113 	rv.restore_scratch = 0;
1114 	rv.need_reload_pte = false;
1115 
1116 	if (check_for_high_segbits) {
1117 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1118 
1119 		if (pgd_reg != -1)
1120 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1121 		else
1122 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1123 
1124 		if (c0_scratch_reg >= 0)
1125 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1126 		else
1127 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1128 
1129 		uasm_i_dsrl_safe(p, scratch, tmp,
1130 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1131 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1132 
1133 		if (pgd_reg == -1) {
1134 			vmalloc_branch_delay_filled = 1;
1135 			/* Clear lower 23 bits of context. */
1136 			uasm_i_dins(p, ptr, 0, 0, 23);
1137 		}
1138 	} else {
1139 		if (pgd_reg != -1)
1140 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1141 		else
1142 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1143 
1144 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1145 
1146 		if (c0_scratch_reg >= 0)
1147 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1148 		else
1149 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1150 
1151 		if (pgd_reg == -1)
1152 			/* Clear lower 23 bits of context. */
1153 			uasm_i_dins(p, ptr, 0, 0, 23);
1154 
1155 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1156 	}
1157 
1158 	if (pgd_reg == -1) {
1159 		vmalloc_branch_delay_filled = 1;
1160 		/* insert bit[63:59] of CAC_BASE into bit[11:6] of ptr */
1161 		uasm_i_ori(p, ptr, ptr, ((u64)(CAC_BASE) >> 53));
1162 
1163 		uasm_i_drotr(p, ptr, ptr, 11);
1164 	}
1165 
1166 #ifdef __PAGETABLE_PMD_FOLDED
1167 #define LOC_PTEP scratch
1168 #else
1169 #define LOC_PTEP ptr
1170 #endif
1171 
1172 	if (!vmalloc_branch_delay_filled)
1173 		/* get pgd offset in bytes */
1174 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1175 
1176 	uasm_l_vmalloc_done(l, *p);
1177 
1178 	/*
1179 	 *			   tmp		ptr
1180 	 * fall-through case =	 badvaddr  *pgd_current
1181 	 * vmalloc case	     =	 badvaddr  swapper_pg_dir
1182 	 */
1183 
1184 	if (vmalloc_branch_delay_filled)
1185 		/* get pgd offset in bytes */
1186 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1187 
1188 #ifdef __PAGETABLE_PMD_FOLDED
1189 	GET_CONTEXT(p, tmp); /* get context reg */
1190 #endif
1191 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1192 
1193 	if (use_lwx_insns()) {
1194 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1195 	} else {
1196 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1197 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1198 	}
1199 
1200 #ifndef __PAGETABLE_PUD_FOLDED
1201 	/* get pud offset in bytes */
1202 	uasm_i_dsrl_safe(p, scratch, tmp, PUD_SHIFT - 3);
1203 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PUD - 1) << 3);
1204 
1205 	if (use_lwx_insns()) {
1206 		UASM_i_LWX(p, ptr, scratch, ptr);
1207 	} else {
1208 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1209 		UASM_i_LW(p, ptr, 0, ptr);
1210 	}
1211 	/* ptr contains a pointer to PMD entry */
1212 	/* tmp contains the address */
1213 #endif
1214 
1215 #ifndef __PAGETABLE_PMD_FOLDED
1216 	/* get pmd offset in bytes */
1217 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1218 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1219 	GET_CONTEXT(p, tmp); /* get context reg */
1220 
1221 	if (use_lwx_insns()) {
1222 		UASM_i_LWX(p, scratch, scratch, ptr);
1223 	} else {
1224 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1225 		UASM_i_LW(p, scratch, 0, ptr);
1226 	}
1227 #endif
1228 	/* Adjust the context during the load latency. */
1229 	build_adjust_context(p, tmp);
1230 
1231 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1232 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1233 	/*
1234 	 * The in the LWX case we don't want to do the load in the
1235 	 * delay slot.	It cannot issue in the same cycle and may be
1236 	 * speculative and unneeded.
1237 	 */
1238 	if (use_lwx_insns())
1239 		uasm_i_nop(p);
1240 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1241 
1242 
1243 	/* build_update_entries */
1244 	if (use_lwx_insns()) {
1245 		even = ptr;
1246 		odd = tmp;
1247 		UASM_i_LWX(p, even, scratch, tmp);
1248 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1249 		UASM_i_LWX(p, odd, scratch, tmp);
1250 	} else {
1251 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1252 		even = tmp;
1253 		odd = ptr;
1254 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1255 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1256 	}
1257 	if (cpu_has_rixi) {
1258 		uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1259 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1260 		uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1261 	} else {
1262 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1263 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1264 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1265 	}
1266 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1267 
1268 	if (c0_scratch_reg >= 0) {
1269 		uasm_i_ehb(p);
1270 		UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1271 		build_tlb_write_entry(p, l, r, tlb_random);
1272 		uasm_l_leave(l, *p);
1273 		rv.restore_scratch = 1;
1274 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1275 		build_tlb_write_entry(p, l, r, tlb_random);
1276 		uasm_l_leave(l, *p);
1277 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1278 	} else {
1279 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1280 		build_tlb_write_entry(p, l, r, tlb_random);
1281 		uasm_l_leave(l, *p);
1282 		rv.restore_scratch = 1;
1283 	}
1284 
1285 	uasm_i_eret(p); /* return from trap */
1286 
1287 	return rv;
1288 }
1289 
1290 /*
1291  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1292  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1293  * slots before the XTLB refill exception handler which belong to the
1294  * unused TLB refill exception.
1295  */
1296 #define MIPS64_REFILL_INSNS 32
1297 
1298 static void build_r4000_tlb_refill_handler(void)
1299 {
1300 	u32 *p = tlb_handler;
1301 	struct uasm_label *l = labels;
1302 	struct uasm_reloc *r = relocs;
1303 	u32 *f;
1304 	unsigned int final_len;
1305 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1306 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1307 
1308 	memset(tlb_handler, 0, sizeof(tlb_handler));
1309 	memset(labels, 0, sizeof(labels));
1310 	memset(relocs, 0, sizeof(relocs));
1311 	memset(final_handler, 0, sizeof(final_handler));
1312 
1313 	if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1314 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1315 							  scratch_reg);
1316 		vmalloc_mode = refill_scratch;
1317 	} else {
1318 		htlb_info.huge_pte = K0;
1319 		htlb_info.restore_scratch = 0;
1320 		htlb_info.need_reload_pte = true;
1321 		vmalloc_mode = refill_noscratch;
1322 		/*
1323 		 * create the plain linear handler
1324 		 */
1325 		if (bcm1250_m3_war()) {
1326 			unsigned int segbits = 44;
1327 
1328 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1329 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1330 			uasm_i_xor(&p, K0, K0, K1);
1331 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1332 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1333 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1334 			uasm_i_or(&p, K0, K0, K1);
1335 			uasm_il_bnez(&p, &r, K0, label_leave);
1336 			/* No need for uasm_i_nop */
1337 		}
1338 
1339 #ifdef CONFIG_64BIT
1340 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1341 #else
1342 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1343 #endif
1344 
1345 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1346 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1347 #endif
1348 
1349 		build_get_ptep(&p, K0, K1);
1350 		build_update_entries(&p, K0, K1);
1351 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1352 		uasm_l_leave(&l, p);
1353 		uasm_i_eret(&p); /* return from trap */
1354 	}
1355 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1356 	uasm_l_tlb_huge_update(&l, p);
1357 	if (htlb_info.need_reload_pte)
1358 		UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1359 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1360 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1361 				   htlb_info.restore_scratch);
1362 #endif
1363 
1364 #ifdef CONFIG_64BIT
1365 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1366 #endif
1367 
1368 	/*
1369 	 * Overflow check: For the 64bit handler, we need at least one
1370 	 * free instruction slot for the wrap-around branch. In worst
1371 	 * case, if the intended insertion point is a delay slot, we
1372 	 * need three, with the second nop'ed and the third being
1373 	 * unused.
1374 	 */
1375 	switch (boot_cpu_type()) {
1376 	default:
1377 		if (sizeof(long) == 4) {
1378 		fallthrough;
1379 	case CPU_LOONGSON2EF:
1380 		/* Loongson2 ebase is different than r4k, we have more space */
1381 			if ((p - tlb_handler) > 64)
1382 				panic("TLB refill handler space exceeded");
1383 			/*
1384 			 * Now fold the handler in the TLB refill handler space.
1385 			 */
1386 			f = final_handler;
1387 			/* Simplest case, just copy the handler. */
1388 			uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1389 			final_len = p - tlb_handler;
1390 			break;
1391 		} else {
1392 			if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1393 			    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1394 				&& uasm_insn_has_bdelay(relocs,
1395 							tlb_handler + MIPS64_REFILL_INSNS - 3)))
1396 				panic("TLB refill handler space exceeded");
1397 			/*
1398 			 * Now fold the handler in the TLB refill handler space.
1399 			 */
1400 			f = final_handler + MIPS64_REFILL_INSNS;
1401 			if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1402 				/* Just copy the handler. */
1403 				uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1404 				final_len = p - tlb_handler;
1405 			} else {
1406 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1407 				const enum label_id ls = label_tlb_huge_update;
1408 #else
1409 				const enum label_id ls = label_vmalloc;
1410 #endif
1411 				u32 *split;
1412 				int ov = 0;
1413 				int i;
1414 
1415 				for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1416 					;
1417 				BUG_ON(i == ARRAY_SIZE(labels));
1418 				split = labels[i].addr;
1419 
1420 				/*
1421 				 * See if we have overflown one way or the other.
1422 				 */
1423 				if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1424 				    split < p - MIPS64_REFILL_INSNS)
1425 					ov = 1;
1426 
1427 				if (ov) {
1428 					/*
1429 					 * Split two instructions before the end.  One
1430 					 * for the branch and one for the instruction
1431 					 * in the delay slot.
1432 					 */
1433 					split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1434 
1435 					/*
1436 					 * If the branch would fall in a delay slot,
1437 					 * we must back up an additional instruction
1438 					 * so that it is no longer in a delay slot.
1439 					 */
1440 					if (uasm_insn_has_bdelay(relocs, split - 1))
1441 						split--;
1442 				}
1443 				/* Copy first part of the handler. */
1444 				uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1445 				f += split - tlb_handler;
1446 
1447 				if (ov) {
1448 					/* Insert branch. */
1449 					uasm_l_split(&l, final_handler);
1450 					uasm_il_b(&f, &r, label_split);
1451 					if (uasm_insn_has_bdelay(relocs, split))
1452 						uasm_i_nop(&f);
1453 					else {
1454 						uasm_copy_handler(relocs, labels,
1455 								  split, split + 1, f);
1456 						uasm_move_labels(labels, f, f + 1, -1);
1457 						f++;
1458 						split++;
1459 					}
1460 				}
1461 
1462 				/* Copy the rest of the handler. */
1463 				uasm_copy_handler(relocs, labels, split, p, final_handler);
1464 				final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1465 					    (p - split);
1466 			}
1467 		}
1468 		break;
1469 	}
1470 
1471 	uasm_resolve_relocs(relocs, labels);
1472 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1473 		 final_len);
1474 
1475 	memcpy((void *)ebase, final_handler, 0x100);
1476 	local_flush_icache_range(ebase, ebase + 0x100);
1477 	dump_handler("r4000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x100));
1478 }
1479 
1480 static void setup_pw(void)
1481 {
1482 	unsigned int pwctl;
1483 	unsigned long pgd_i, pgd_w;
1484 #ifndef __PAGETABLE_PMD_FOLDED
1485 	unsigned long pmd_i, pmd_w;
1486 #endif
1487 	unsigned long pt_i, pt_w;
1488 	unsigned long pte_i, pte_w;
1489 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1490 	unsigned long psn;
1491 
1492 	psn = ilog2(_PAGE_HUGE);     /* bit used to indicate huge page */
1493 #endif
1494 	pgd_i = PGDIR_SHIFT;  /* 1st level PGD */
1495 #ifndef __PAGETABLE_PMD_FOLDED
1496 	pgd_w = PGDIR_SHIFT - PMD_SHIFT + PGD_ORDER;
1497 
1498 	pmd_i = PMD_SHIFT;    /* 2nd level PMD */
1499 	pmd_w = PMD_SHIFT - PAGE_SHIFT;
1500 #else
1501 	pgd_w = PGDIR_SHIFT - PAGE_SHIFT + PGD_ORDER;
1502 #endif
1503 
1504 	pt_i  = PAGE_SHIFT;    /* 3rd level PTE */
1505 	pt_w  = PAGE_SHIFT - 3;
1506 
1507 	pte_i = ilog2(_PAGE_GLOBAL);
1508 	pte_w = 0;
1509 	pwctl = 1 << 30; /* Set PWDirExt */
1510 
1511 #ifndef __PAGETABLE_PMD_FOLDED
1512 	write_c0_pwfield(pgd_i << 24 | pmd_i << 12 | pt_i << 6 | pte_i);
1513 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pmd_w << 12 | pt_w << 6 | pte_w);
1514 #else
1515 	write_c0_pwfield(pgd_i << 24 | pt_i << 6 | pte_i);
1516 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pt_w << 6 | pte_w);
1517 #endif
1518 
1519 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1520 	pwctl |= (1 << 6 | psn);
1521 #endif
1522 	write_c0_pwctl(pwctl);
1523 	write_c0_kpgd((long)swapper_pg_dir);
1524 	kscratch_used_mask |= (1 << 7); /* KScratch6 is used for KPGD */
1525 }
1526 
1527 static void build_loongson3_tlb_refill_handler(void)
1528 {
1529 	u32 *p = tlb_handler;
1530 	struct uasm_label *l = labels;
1531 	struct uasm_reloc *r = relocs;
1532 
1533 	memset(labels, 0, sizeof(labels));
1534 	memset(relocs, 0, sizeof(relocs));
1535 	memset(tlb_handler, 0, sizeof(tlb_handler));
1536 
1537 	if (check_for_high_segbits) {
1538 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1539 		uasm_i_dsrl_safe(&p, K1, K0, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1540 		uasm_il_beqz(&p, &r, K1, label_vmalloc);
1541 		uasm_i_nop(&p);
1542 
1543 		uasm_il_bgez(&p, &r, K0, label_large_segbits_fault);
1544 		uasm_i_nop(&p);
1545 		uasm_l_vmalloc(&l, p);
1546 	}
1547 
1548 	uasm_i_dmfc0(&p, K1, C0_PGD);
1549 
1550 	uasm_i_lddir(&p, K0, K1, 3);  /* global page dir */
1551 #ifndef __PAGETABLE_PMD_FOLDED
1552 	uasm_i_lddir(&p, K1, K0, 1);  /* middle page dir */
1553 #endif
1554 	uasm_i_ldpte(&p, K1, 0);      /* even */
1555 	uasm_i_ldpte(&p, K1, 1);      /* odd */
1556 	uasm_i_tlbwr(&p);
1557 
1558 	/* restore page mask */
1559 	if (PM_DEFAULT_MASK >> 16) {
1560 		uasm_i_lui(&p, K0, PM_DEFAULT_MASK >> 16);
1561 		uasm_i_ori(&p, K0, K0, PM_DEFAULT_MASK & 0xffff);
1562 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1563 	} else if (PM_DEFAULT_MASK) {
1564 		uasm_i_ori(&p, K0, 0, PM_DEFAULT_MASK);
1565 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1566 	} else {
1567 		uasm_i_mtc0(&p, 0, C0_PAGEMASK);
1568 	}
1569 
1570 	uasm_i_eret(&p);
1571 
1572 	if (check_for_high_segbits) {
1573 		uasm_l_large_segbits_fault(&l, p);
1574 		UASM_i_LA(&p, K1, (unsigned long)tlb_do_page_fault_0);
1575 		uasm_i_jr(&p, K1);
1576 		uasm_i_nop(&p);
1577 	}
1578 
1579 	uasm_resolve_relocs(relocs, labels);
1580 	memcpy((void *)(ebase + 0x80), tlb_handler, 0x80);
1581 	local_flush_icache_range(ebase + 0x80, ebase + 0x100);
1582 	dump_handler("loongson3_tlb_refill",
1583 		     (u32 *)(ebase + 0x80), (u32 *)(ebase + 0x100));
1584 }
1585 
1586 static void build_setup_pgd(void)
1587 {
1588 	const int a0 = 4;
1589 	const int __maybe_unused a1 = 5;
1590 	const int __maybe_unused a2 = 6;
1591 	u32 *p = (u32 *)msk_isa16_mode((ulong)tlbmiss_handler_setup_pgd);
1592 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1593 	long pgdc = (long)pgd_current;
1594 #endif
1595 
1596 	memset(p, 0, tlbmiss_handler_setup_pgd_end - (char *)p);
1597 	memset(labels, 0, sizeof(labels));
1598 	memset(relocs, 0, sizeof(relocs));
1599 	pgd_reg = allocate_kscratch();
1600 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1601 	if (pgd_reg == -1) {
1602 		struct uasm_label *l = labels;
1603 		struct uasm_reloc *r = relocs;
1604 
1605 		/* PGD << 11 in c0_Context */
1606 		/*
1607 		 * If it is a ckseg0 address, convert to a physical
1608 		 * address.  Shifting right by 29 and adding 4 will
1609 		 * result in zero for these addresses.
1610 		 *
1611 		 */
1612 		UASM_i_SRA(&p, a1, a0, 29);
1613 		UASM_i_ADDIU(&p, a1, a1, 4);
1614 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1615 		uasm_i_nop(&p);
1616 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1617 		uasm_l_tlbl_goaround1(&l, p);
1618 		UASM_i_SLL(&p, a0, a0, 11);
1619 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1620 		uasm_i_jr(&p, 31);
1621 		uasm_i_ehb(&p);
1622 	} else {
1623 		/* PGD in c0_KScratch */
1624 		if (cpu_has_ldpte)
1625 			UASM_i_MTC0(&p, a0, C0_PWBASE);
1626 		else
1627 			UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1628 		uasm_i_jr(&p, 31);
1629 		uasm_i_ehb(&p);
1630 	}
1631 #else
1632 #ifdef CONFIG_SMP
1633 	/* Save PGD to pgd_current[smp_processor_id()] */
1634 	UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1635 	UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1636 	UASM_i_LA_mostly(&p, a2, pgdc);
1637 	UASM_i_ADDU(&p, a2, a2, a1);
1638 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1639 #else
1640 	UASM_i_LA_mostly(&p, a2, pgdc);
1641 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1642 #endif /* SMP */
1643 
1644 	/* if pgd_reg is allocated, save PGD also to scratch register */
1645 	if (pgd_reg != -1) {
1646 		UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1647 		uasm_i_jr(&p, 31);
1648 		uasm_i_ehb(&p);
1649 	} else {
1650 		uasm_i_jr(&p, 31);
1651 		uasm_i_nop(&p);
1652 	}
1653 #endif
1654 	if (p >= (u32 *)tlbmiss_handler_setup_pgd_end)
1655 		panic("tlbmiss_handler_setup_pgd space exceeded");
1656 
1657 	uasm_resolve_relocs(relocs, labels);
1658 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1659 		 (unsigned int)(p - (u32 *)tlbmiss_handler_setup_pgd));
1660 
1661 	dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1662 					tlbmiss_handler_setup_pgd_end);
1663 }
1664 
1665 static void
1666 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1667 {
1668 #ifdef CONFIG_SMP
1669 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
1670 		uasm_i_sync(p, 0);
1671 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1672 	if (cpu_has_64bits)
1673 		uasm_i_lld(p, pte, 0, ptr);
1674 	else
1675 # endif
1676 		UASM_i_LL(p, pte, 0, ptr);
1677 #else
1678 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1679 	if (cpu_has_64bits)
1680 		uasm_i_ld(p, pte, 0, ptr);
1681 	else
1682 # endif
1683 		UASM_i_LW(p, pte, 0, ptr);
1684 #endif
1685 }
1686 
1687 static void
1688 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1689 	unsigned int mode, unsigned int scratch)
1690 {
1691 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1692 	unsigned int swmode = mode & ~hwmode;
1693 
1694 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_64bits) {
1695 		uasm_i_lui(p, scratch, swmode >> 16);
1696 		uasm_i_or(p, pte, pte, scratch);
1697 		BUG_ON(swmode & 0xffff);
1698 	} else {
1699 		uasm_i_ori(p, pte, pte, mode);
1700 	}
1701 
1702 #ifdef CONFIG_SMP
1703 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1704 	if (cpu_has_64bits)
1705 		uasm_i_scd(p, pte, 0, ptr);
1706 	else
1707 # endif
1708 		UASM_i_SC(p, pte, 0, ptr);
1709 
1710 	if (r10000_llsc_war())
1711 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1712 	else
1713 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1714 
1715 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1716 	if (!cpu_has_64bits) {
1717 		/* no uasm_i_nop needed */
1718 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1719 		uasm_i_ori(p, pte, pte, hwmode);
1720 		BUG_ON(hwmode & ~0xffff);
1721 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1722 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1723 		/* no uasm_i_nop needed */
1724 		uasm_i_lw(p, pte, 0, ptr);
1725 	} else
1726 		uasm_i_nop(p);
1727 # else
1728 	uasm_i_nop(p);
1729 # endif
1730 #else
1731 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1732 	if (cpu_has_64bits)
1733 		uasm_i_sd(p, pte, 0, ptr);
1734 	else
1735 # endif
1736 		UASM_i_SW(p, pte, 0, ptr);
1737 
1738 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1739 	if (!cpu_has_64bits) {
1740 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1741 		uasm_i_ori(p, pte, pte, hwmode);
1742 		BUG_ON(hwmode & ~0xffff);
1743 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1744 		uasm_i_lw(p, pte, 0, ptr);
1745 	}
1746 # endif
1747 #endif
1748 }
1749 
1750 /*
1751  * Check if PTE is present, if not then jump to LABEL. PTR points to
1752  * the page table where this PTE is located, PTE will be re-loaded
1753  * with it's original value.
1754  */
1755 static void
1756 build_pte_present(u32 **p, struct uasm_reloc **r,
1757 		  int pte, int ptr, int scratch, enum label_id lid)
1758 {
1759 	int t = scratch >= 0 ? scratch : pte;
1760 	int cur = pte;
1761 
1762 	if (cpu_has_rixi) {
1763 		if (use_bbit_insns()) {
1764 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1765 			uasm_i_nop(p);
1766 		} else {
1767 			if (_PAGE_PRESENT_SHIFT) {
1768 				uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1769 				cur = t;
1770 			}
1771 			uasm_i_andi(p, t, cur, 1);
1772 			uasm_il_beqz(p, r, t, lid);
1773 			if (pte == t)
1774 				/* You lose the SMP race :-(*/
1775 				iPTE_LW(p, pte, ptr);
1776 		}
1777 	} else {
1778 		if (_PAGE_PRESENT_SHIFT) {
1779 			uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1780 			cur = t;
1781 		}
1782 		uasm_i_andi(p, t, cur,
1783 			(_PAGE_PRESENT | _PAGE_NO_READ) >> _PAGE_PRESENT_SHIFT);
1784 		uasm_i_xori(p, t, t, _PAGE_PRESENT >> _PAGE_PRESENT_SHIFT);
1785 		uasm_il_bnez(p, r, t, lid);
1786 		if (pte == t)
1787 			/* You lose the SMP race :-(*/
1788 			iPTE_LW(p, pte, ptr);
1789 	}
1790 }
1791 
1792 /* Make PTE valid, store result in PTR. */
1793 static void
1794 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1795 		 unsigned int ptr, unsigned int scratch)
1796 {
1797 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1798 
1799 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1800 }
1801 
1802 /*
1803  * Check if PTE can be written to, if not branch to LABEL. Regardless
1804  * restore PTE with value from PTR when done.
1805  */
1806 static void
1807 build_pte_writable(u32 **p, struct uasm_reloc **r,
1808 		   unsigned int pte, unsigned int ptr, int scratch,
1809 		   enum label_id lid)
1810 {
1811 	int t = scratch >= 0 ? scratch : pte;
1812 	int cur = pte;
1813 
1814 	if (_PAGE_PRESENT_SHIFT) {
1815 		uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1816 		cur = t;
1817 	}
1818 	uasm_i_andi(p, t, cur,
1819 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1820 	uasm_i_xori(p, t, t,
1821 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1822 	uasm_il_bnez(p, r, t, lid);
1823 	if (pte == t)
1824 		/* You lose the SMP race :-(*/
1825 		iPTE_LW(p, pte, ptr);
1826 	else
1827 		uasm_i_nop(p);
1828 }
1829 
1830 /* Make PTE writable, update software status bits as well, then store
1831  * at PTR.
1832  */
1833 static void
1834 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1835 		 unsigned int ptr, unsigned int scratch)
1836 {
1837 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1838 			     | _PAGE_DIRTY);
1839 
1840 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1841 }
1842 
1843 /*
1844  * Check if PTE can be modified, if not branch to LABEL. Regardless
1845  * restore PTE with value from PTR when done.
1846  */
1847 static void
1848 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1849 		     unsigned int pte, unsigned int ptr, int scratch,
1850 		     enum label_id lid)
1851 {
1852 	if (use_bbit_insns()) {
1853 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1854 		uasm_i_nop(p);
1855 	} else {
1856 		int t = scratch >= 0 ? scratch : pte;
1857 		uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1858 		uasm_i_andi(p, t, t, 1);
1859 		uasm_il_beqz(p, r, t, lid);
1860 		if (pte == t)
1861 			/* You lose the SMP race :-(*/
1862 			iPTE_LW(p, pte, ptr);
1863 	}
1864 }
1865 
1866 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1867 
1868 
1869 /*
1870  * R3000 style TLB load/store/modify handlers.
1871  */
1872 
1873 /*
1874  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1875  * Then it returns.
1876  */
1877 static void
1878 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1879 {
1880 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1881 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1882 	uasm_i_tlbwi(p);
1883 	uasm_i_jr(p, tmp);
1884 	uasm_i_rfe(p); /* branch delay */
1885 }
1886 
1887 /*
1888  * This places the pte into ENTRYLO0 and writes it with tlbwi
1889  * or tlbwr as appropriate.  This is because the index register
1890  * may have the probe fail bit set as a result of a trap on a
1891  * kseg2 access, i.e. without refill.  Then it returns.
1892  */
1893 static void
1894 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1895 			     struct uasm_reloc **r, unsigned int pte,
1896 			     unsigned int tmp)
1897 {
1898 	uasm_i_mfc0(p, tmp, C0_INDEX);
1899 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1900 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1901 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1902 	uasm_i_tlbwi(p); /* cp0 delay */
1903 	uasm_i_jr(p, tmp);
1904 	uasm_i_rfe(p); /* branch delay */
1905 	uasm_l_r3000_write_probe_fail(l, *p);
1906 	uasm_i_tlbwr(p); /* cp0 delay */
1907 	uasm_i_jr(p, tmp);
1908 	uasm_i_rfe(p); /* branch delay */
1909 }
1910 
1911 static void
1912 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1913 				   unsigned int ptr)
1914 {
1915 	long pgdc = (long)pgd_current;
1916 
1917 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1918 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1919 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1920 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1921 	uasm_i_sll(p, pte, pte, 2);
1922 	uasm_i_addu(p, ptr, ptr, pte);
1923 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1924 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1925 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1926 	uasm_i_addu(p, ptr, ptr, pte);
1927 	uasm_i_lw(p, pte, 0, ptr);
1928 	uasm_i_tlbp(p); /* load delay */
1929 }
1930 
1931 static void build_r3000_tlb_load_handler(void)
1932 {
1933 	u32 *p = (u32 *)handle_tlbl;
1934 	struct uasm_label *l = labels;
1935 	struct uasm_reloc *r = relocs;
1936 
1937 	memset(p, 0, handle_tlbl_end - (char *)p);
1938 	memset(labels, 0, sizeof(labels));
1939 	memset(relocs, 0, sizeof(relocs));
1940 
1941 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1942 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1943 	uasm_i_nop(&p); /* load delay */
1944 	build_make_valid(&p, &r, K0, K1, -1);
1945 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1946 
1947 	uasm_l_nopage_tlbl(&l, p);
1948 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1949 	uasm_i_nop(&p);
1950 
1951 	if (p >= (u32 *)handle_tlbl_end)
1952 		panic("TLB load handler fastpath space exceeded");
1953 
1954 	uasm_resolve_relocs(relocs, labels);
1955 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1956 		 (unsigned int)(p - (u32 *)handle_tlbl));
1957 
1958 	dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_end);
1959 }
1960 
1961 static void build_r3000_tlb_store_handler(void)
1962 {
1963 	u32 *p = (u32 *)handle_tlbs;
1964 	struct uasm_label *l = labels;
1965 	struct uasm_reloc *r = relocs;
1966 
1967 	memset(p, 0, handle_tlbs_end - (char *)p);
1968 	memset(labels, 0, sizeof(labels));
1969 	memset(relocs, 0, sizeof(relocs));
1970 
1971 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1972 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1973 	uasm_i_nop(&p); /* load delay */
1974 	build_make_write(&p, &r, K0, K1, -1);
1975 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1976 
1977 	uasm_l_nopage_tlbs(&l, p);
1978 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1979 	uasm_i_nop(&p);
1980 
1981 	if (p >= (u32 *)handle_tlbs_end)
1982 		panic("TLB store handler fastpath space exceeded");
1983 
1984 	uasm_resolve_relocs(relocs, labels);
1985 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1986 		 (unsigned int)(p - (u32 *)handle_tlbs));
1987 
1988 	dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_end);
1989 }
1990 
1991 static void build_r3000_tlb_modify_handler(void)
1992 {
1993 	u32 *p = (u32 *)handle_tlbm;
1994 	struct uasm_label *l = labels;
1995 	struct uasm_reloc *r = relocs;
1996 
1997 	memset(p, 0, handle_tlbm_end - (char *)p);
1998 	memset(labels, 0, sizeof(labels));
1999 	memset(relocs, 0, sizeof(relocs));
2000 
2001 	build_r3000_tlbchange_handler_head(&p, K0, K1);
2002 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
2003 	uasm_i_nop(&p); /* load delay */
2004 	build_make_write(&p, &r, K0, K1, -1);
2005 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
2006 
2007 	uasm_l_nopage_tlbm(&l, p);
2008 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2009 	uasm_i_nop(&p);
2010 
2011 	if (p >= (u32 *)handle_tlbm_end)
2012 		panic("TLB modify handler fastpath space exceeded");
2013 
2014 	uasm_resolve_relocs(relocs, labels);
2015 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2016 		 (unsigned int)(p - (u32 *)handle_tlbm));
2017 
2018 	dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_end);
2019 }
2020 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
2021 
2022 static bool cpu_has_tlbex_tlbp_race(void)
2023 {
2024 	/*
2025 	 * When a Hardware Table Walker is running it can replace TLB entries
2026 	 * at any time, leading to a race between it & the CPU.
2027 	 */
2028 	if (cpu_has_htw)
2029 		return true;
2030 
2031 	/*
2032 	 * If the CPU shares FTLB RAM with its siblings then our entry may be
2033 	 * replaced at any time by a sibling performing a write to the FTLB.
2034 	 */
2035 	if (cpu_has_shared_ftlb_ram)
2036 		return true;
2037 
2038 	/* In all other cases there ought to be no race condition to handle */
2039 	return false;
2040 }
2041 
2042 /*
2043  * R4000 style TLB load/store/modify handlers.
2044  */
2045 static struct work_registers
2046 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
2047 				   struct uasm_reloc **r)
2048 {
2049 	struct work_registers wr = build_get_work_registers(p);
2050 
2051 #ifdef CONFIG_64BIT
2052 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
2053 #else
2054 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
2055 #endif
2056 
2057 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2058 	/*
2059 	 * For huge tlb entries, pmd doesn't contain an address but
2060 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
2061 	 * see if we need to jump to huge tlb processing.
2062 	 */
2063 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
2064 #endif
2065 
2066 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
2067 	UASM_i_LW(p, wr.r2, 0, wr.r2);
2068 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
2069 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
2070 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
2071 
2072 #ifdef CONFIG_SMP
2073 	uasm_l_smp_pgtable_change(l, *p);
2074 #endif
2075 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
2076 	if (!m4kc_tlbp_war()) {
2077 		build_tlb_probe_entry(p);
2078 		if (cpu_has_tlbex_tlbp_race()) {
2079 			/* race condition happens, leaving */
2080 			uasm_i_ehb(p);
2081 			uasm_i_mfc0(p, wr.r3, C0_INDEX);
2082 			uasm_il_bltz(p, r, wr.r3, label_leave);
2083 			uasm_i_nop(p);
2084 		}
2085 	}
2086 	return wr;
2087 }
2088 
2089 static void
2090 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
2091 				   struct uasm_reloc **r, unsigned int tmp,
2092 				   unsigned int ptr)
2093 {
2094 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
2095 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
2096 	build_update_entries(p, tmp, ptr);
2097 	build_tlb_write_entry(p, l, r, tlb_indexed);
2098 	uasm_l_leave(l, *p);
2099 	build_restore_work_registers(p);
2100 	uasm_i_eret(p); /* return from trap */
2101 
2102 #ifdef CONFIG_64BIT
2103 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
2104 #endif
2105 }
2106 
2107 static void build_r4000_tlb_load_handler(void)
2108 {
2109 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbl);
2110 	struct uasm_label *l = labels;
2111 	struct uasm_reloc *r = relocs;
2112 	struct work_registers wr;
2113 
2114 	memset(p, 0, handle_tlbl_end - (char *)p);
2115 	memset(labels, 0, sizeof(labels));
2116 	memset(relocs, 0, sizeof(relocs));
2117 
2118 	if (bcm1250_m3_war()) {
2119 		unsigned int segbits = 44;
2120 
2121 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
2122 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
2123 		uasm_i_xor(&p, K0, K0, K1);
2124 		uasm_i_dsrl_safe(&p, K1, K0, 62);
2125 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
2126 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
2127 		uasm_i_or(&p, K0, K0, K1);
2128 		uasm_il_bnez(&p, &r, K0, label_leave);
2129 		/* No need for uasm_i_nop */
2130 	}
2131 
2132 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2133 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2134 	if (m4kc_tlbp_war())
2135 		build_tlb_probe_entry(&p);
2136 
2137 	if (cpu_has_rixi && !cpu_has_rixiex) {
2138 		/*
2139 		 * If the page is not _PAGE_VALID, RI or XI could not
2140 		 * have triggered it.  Skip the expensive test..
2141 		 */
2142 		if (use_bbit_insns()) {
2143 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2144 				      label_tlbl_goaround1);
2145 		} else {
2146 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2147 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
2148 		}
2149 		uasm_i_nop(&p);
2150 
2151 		/*
2152 		 * Warn if something may race with us & replace the TLB entry
2153 		 * before we read it here. Everything with such races should
2154 		 * also have dedicated RiXi exception handlers, so this
2155 		 * shouldn't be hit.
2156 		 */
2157 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2158 
2159 		uasm_i_tlbr(&p);
2160 
2161 		switch (current_cpu_type()) {
2162 		case CPU_CAVIUM_OCTEON:
2163 		case CPU_CAVIUM_OCTEON_PLUS:
2164 		case CPU_CAVIUM_OCTEON2:
2165 			break;
2166 		default:
2167 			if (cpu_has_mips_r2_exec_hazard)
2168 				uasm_i_ehb(&p);
2169 			break;
2170 		}
2171 
2172 		/* Examine  entrylo 0 or 1 based on ptr. */
2173 		if (use_bbit_insns()) {
2174 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2175 		} else {
2176 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2177 			uasm_i_beqz(&p, wr.r3, 8);
2178 		}
2179 		/* load it in the delay slot*/
2180 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2181 		/* load it if ptr is odd */
2182 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2183 		/*
2184 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2185 		 * XI must have triggered it.
2186 		 */
2187 		if (use_bbit_insns()) {
2188 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2189 			uasm_i_nop(&p);
2190 			uasm_l_tlbl_goaround1(&l, p);
2191 		} else {
2192 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2193 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2194 			uasm_i_nop(&p);
2195 		}
2196 		uasm_l_tlbl_goaround1(&l, p);
2197 	}
2198 	build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2199 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2200 
2201 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2202 	/*
2203 	 * This is the entry point when build_r4000_tlbchange_handler_head
2204 	 * spots a huge page.
2205 	 */
2206 	uasm_l_tlb_huge_update(&l, p);
2207 	iPTE_LW(&p, wr.r1, wr.r2);
2208 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2209 	build_tlb_probe_entry(&p);
2210 
2211 	if (cpu_has_rixi && !cpu_has_rixiex) {
2212 		/*
2213 		 * If the page is not _PAGE_VALID, RI or XI could not
2214 		 * have triggered it.  Skip the expensive test..
2215 		 */
2216 		if (use_bbit_insns()) {
2217 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2218 				      label_tlbl_goaround2);
2219 		} else {
2220 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2221 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2222 		}
2223 		uasm_i_nop(&p);
2224 
2225 		/*
2226 		 * Warn if something may race with us & replace the TLB entry
2227 		 * before we read it here. Everything with such races should
2228 		 * also have dedicated RiXi exception handlers, so this
2229 		 * shouldn't be hit.
2230 		 */
2231 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2232 
2233 		uasm_i_tlbr(&p);
2234 
2235 		switch (current_cpu_type()) {
2236 		case CPU_CAVIUM_OCTEON:
2237 		case CPU_CAVIUM_OCTEON_PLUS:
2238 		case CPU_CAVIUM_OCTEON2:
2239 			break;
2240 		default:
2241 			if (cpu_has_mips_r2_exec_hazard)
2242 				uasm_i_ehb(&p);
2243 			break;
2244 		}
2245 
2246 		/* Examine  entrylo 0 or 1 based on ptr. */
2247 		if (use_bbit_insns()) {
2248 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2249 		} else {
2250 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2251 			uasm_i_beqz(&p, wr.r3, 8);
2252 		}
2253 		/* load it in the delay slot*/
2254 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2255 		/* load it if ptr is odd */
2256 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2257 		/*
2258 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2259 		 * XI must have triggered it.
2260 		 */
2261 		if (use_bbit_insns()) {
2262 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2263 		} else {
2264 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2265 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2266 		}
2267 		if (PM_DEFAULT_MASK == 0)
2268 			uasm_i_nop(&p);
2269 		/*
2270 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
2271 		 * it is restored in build_huge_tlb_write_entry.
2272 		 */
2273 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2274 
2275 		uasm_l_tlbl_goaround2(&l, p);
2276 	}
2277 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2278 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2279 #endif
2280 
2281 	uasm_l_nopage_tlbl(&l, p);
2282 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2283 		uasm_i_sync(&p, 0);
2284 	build_restore_work_registers(&p);
2285 #ifdef CONFIG_CPU_MICROMIPS
2286 	if ((unsigned long)tlb_do_page_fault_0 & 1) {
2287 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2288 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2289 		uasm_i_jr(&p, K0);
2290 	} else
2291 #endif
2292 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2293 	uasm_i_nop(&p);
2294 
2295 	if (p >= (u32 *)handle_tlbl_end)
2296 		panic("TLB load handler fastpath space exceeded");
2297 
2298 	uasm_resolve_relocs(relocs, labels);
2299 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2300 		 (unsigned int)(p - (u32 *)handle_tlbl));
2301 
2302 	dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_end);
2303 }
2304 
2305 static void build_r4000_tlb_store_handler(void)
2306 {
2307 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbs);
2308 	struct uasm_label *l = labels;
2309 	struct uasm_reloc *r = relocs;
2310 	struct work_registers wr;
2311 
2312 	memset(p, 0, handle_tlbs_end - (char *)p);
2313 	memset(labels, 0, sizeof(labels));
2314 	memset(relocs, 0, sizeof(relocs));
2315 
2316 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2317 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2318 	if (m4kc_tlbp_war())
2319 		build_tlb_probe_entry(&p);
2320 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2321 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2322 
2323 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2324 	/*
2325 	 * This is the entry point when
2326 	 * build_r4000_tlbchange_handler_head spots a huge page.
2327 	 */
2328 	uasm_l_tlb_huge_update(&l, p);
2329 	iPTE_LW(&p, wr.r1, wr.r2);
2330 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2331 	build_tlb_probe_entry(&p);
2332 	uasm_i_ori(&p, wr.r1, wr.r1,
2333 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2334 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2335 #endif
2336 
2337 	uasm_l_nopage_tlbs(&l, p);
2338 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2339 		uasm_i_sync(&p, 0);
2340 	build_restore_work_registers(&p);
2341 #ifdef CONFIG_CPU_MICROMIPS
2342 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2343 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2344 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2345 		uasm_i_jr(&p, K0);
2346 	} else
2347 #endif
2348 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2349 	uasm_i_nop(&p);
2350 
2351 	if (p >= (u32 *)handle_tlbs_end)
2352 		panic("TLB store handler fastpath space exceeded");
2353 
2354 	uasm_resolve_relocs(relocs, labels);
2355 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2356 		 (unsigned int)(p - (u32 *)handle_tlbs));
2357 
2358 	dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_end);
2359 }
2360 
2361 static void build_r4000_tlb_modify_handler(void)
2362 {
2363 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbm);
2364 	struct uasm_label *l = labels;
2365 	struct uasm_reloc *r = relocs;
2366 	struct work_registers wr;
2367 
2368 	memset(p, 0, handle_tlbm_end - (char *)p);
2369 	memset(labels, 0, sizeof(labels));
2370 	memset(relocs, 0, sizeof(relocs));
2371 
2372 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2373 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2374 	if (m4kc_tlbp_war())
2375 		build_tlb_probe_entry(&p);
2376 	/* Present and writable bits set, set accessed and dirty bits. */
2377 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2378 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2379 
2380 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2381 	/*
2382 	 * This is the entry point when
2383 	 * build_r4000_tlbchange_handler_head spots a huge page.
2384 	 */
2385 	uasm_l_tlb_huge_update(&l, p);
2386 	iPTE_LW(&p, wr.r1, wr.r2);
2387 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2388 	build_tlb_probe_entry(&p);
2389 	uasm_i_ori(&p, wr.r1, wr.r1,
2390 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2391 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 0);
2392 #endif
2393 
2394 	uasm_l_nopage_tlbm(&l, p);
2395 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2396 		uasm_i_sync(&p, 0);
2397 	build_restore_work_registers(&p);
2398 #ifdef CONFIG_CPU_MICROMIPS
2399 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2400 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2401 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2402 		uasm_i_jr(&p, K0);
2403 	} else
2404 #endif
2405 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2406 	uasm_i_nop(&p);
2407 
2408 	if (p >= (u32 *)handle_tlbm_end)
2409 		panic("TLB modify handler fastpath space exceeded");
2410 
2411 	uasm_resolve_relocs(relocs, labels);
2412 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2413 		 (unsigned int)(p - (u32 *)handle_tlbm));
2414 
2415 	dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_end);
2416 }
2417 
2418 static void flush_tlb_handlers(void)
2419 {
2420 	local_flush_icache_range((unsigned long)handle_tlbl,
2421 			   (unsigned long)handle_tlbl_end);
2422 	local_flush_icache_range((unsigned long)handle_tlbs,
2423 			   (unsigned long)handle_tlbs_end);
2424 	local_flush_icache_range((unsigned long)handle_tlbm,
2425 			   (unsigned long)handle_tlbm_end);
2426 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2427 			   (unsigned long)tlbmiss_handler_setup_pgd_end);
2428 }
2429 
2430 static void print_htw_config(void)
2431 {
2432 	unsigned long config;
2433 	unsigned int pwctl;
2434 	const int field = 2 * sizeof(unsigned long);
2435 
2436 	config = read_c0_pwfield();
2437 	pr_debug("PWField (0x%0*lx): GDI: 0x%02lx  UDI: 0x%02lx  MDI: 0x%02lx  PTI: 0x%02lx  PTEI: 0x%02lx\n",
2438 		field, config,
2439 		(config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2440 		(config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2441 		(config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2442 		(config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2443 		(config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2444 
2445 	config = read_c0_pwsize();
2446 	pr_debug("PWSize  (0x%0*lx): PS: 0x%lx  GDW: 0x%02lx  UDW: 0x%02lx  MDW: 0x%02lx  PTW: 0x%02lx  PTEW: 0x%02lx\n",
2447 		field, config,
2448 		(config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2449 		(config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2450 		(config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2451 		(config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2452 		(config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2453 		(config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2454 
2455 	pwctl = read_c0_pwctl();
2456 	pr_debug("PWCtl   (0x%x): PWEn: 0x%x  XK: 0x%x  XS: 0x%x  XU: 0x%x  DPH: 0x%x  HugePg: 0x%x  Psn: 0x%x\n",
2457 		pwctl,
2458 		(pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2459 		(pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2460 		(pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2461 		(pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2462 		(pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2463 		(pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2464 		(pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2465 }
2466 
2467 static void config_htw_params(void)
2468 {
2469 	unsigned long pwfield, pwsize, ptei;
2470 	unsigned int config;
2471 
2472 	/*
2473 	 * We are using 2-level page tables, so we only need to
2474 	 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2475 	 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2476 	 * write values less than 0xc in these fields because the entire
2477 	 * write will be dropped. As a result of which, we must preserve
2478 	 * the original reset values and overwrite only what we really want.
2479 	 */
2480 
2481 	pwfield = read_c0_pwfield();
2482 	/* re-initialize the GDI field */
2483 	pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2484 	pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2485 	/* re-initialize the PTI field including the even/odd bit */
2486 	pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2487 	pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2488 	if (CONFIG_PGTABLE_LEVELS >= 3) {
2489 		pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2490 		pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2491 	}
2492 	/* Set the PTEI right shift */
2493 	ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2494 	pwfield |= ptei;
2495 	write_c0_pwfield(pwfield);
2496 	/* Check whether the PTEI value is supported */
2497 	back_to_back_c0_hazard();
2498 	pwfield = read_c0_pwfield();
2499 	if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2500 		!= ptei) {
2501 		pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2502 			ptei);
2503 		/*
2504 		 * Drop option to avoid HTW being enabled via another path
2505 		 * (eg htw_reset())
2506 		 */
2507 		current_cpu_data.options &= ~MIPS_CPU_HTW;
2508 		return;
2509 	}
2510 
2511 	pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2512 	pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2513 	if (CONFIG_PGTABLE_LEVELS >= 3)
2514 		pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2515 
2516 	/* Set pointer size to size of directory pointers */
2517 	if (IS_ENABLED(CONFIG_64BIT))
2518 		pwsize |= MIPS_PWSIZE_PS_MASK;
2519 	/* PTEs may be multiple pointers long (e.g. with XPA) */
2520 	pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2521 			& MIPS_PWSIZE_PTEW_MASK;
2522 
2523 	write_c0_pwsize(pwsize);
2524 
2525 	/* Make sure everything is set before we enable the HTW */
2526 	back_to_back_c0_hazard();
2527 
2528 	/*
2529 	 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2530 	 * the pwctl fields.
2531 	 */
2532 	config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2533 	if (IS_ENABLED(CONFIG_64BIT))
2534 		config |= MIPS_PWCTL_XU_MASK;
2535 	write_c0_pwctl(config);
2536 	pr_info("Hardware Page Table Walker enabled\n");
2537 
2538 	print_htw_config();
2539 }
2540 
2541 static void config_xpa_params(void)
2542 {
2543 #ifdef CONFIG_XPA
2544 	unsigned int pagegrain;
2545 
2546 	if (mips_xpa_disabled) {
2547 		pr_info("Extended Physical Addressing (XPA) disabled\n");
2548 		return;
2549 	}
2550 
2551 	pagegrain = read_c0_pagegrain();
2552 	write_c0_pagegrain(pagegrain | PG_ELPA);
2553 	back_to_back_c0_hazard();
2554 	pagegrain = read_c0_pagegrain();
2555 
2556 	if (pagegrain & PG_ELPA)
2557 		pr_info("Extended Physical Addressing (XPA) enabled\n");
2558 	else
2559 		panic("Extended Physical Addressing (XPA) disabled");
2560 #endif
2561 }
2562 
2563 static void check_pabits(void)
2564 {
2565 	unsigned long entry;
2566 	unsigned pabits, fillbits;
2567 
2568 	if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2569 		/*
2570 		 * We'll only be making use of the fact that we can rotate bits
2571 		 * into the fill if the CPU supports RIXI, so don't bother
2572 		 * probing this for CPUs which don't.
2573 		 */
2574 		return;
2575 	}
2576 
2577 	write_c0_entrylo0(~0ul);
2578 	back_to_back_c0_hazard();
2579 	entry = read_c0_entrylo0();
2580 
2581 	/* clear all non-PFN bits */
2582 	entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2583 	entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2584 
2585 	/* find a lower bound on PABITS, and upper bound on fill bits */
2586 	pabits = fls_long(entry) + 6;
2587 	fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2588 
2589 	/* minus the RI & XI bits */
2590 	fillbits -= min_t(unsigned, fillbits, 2);
2591 
2592 	if (fillbits >= ilog2(_PAGE_NO_EXEC))
2593 		fill_includes_sw_bits = true;
2594 
2595 	pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2596 }
2597 
2598 void build_tlb_refill_handler(void)
2599 {
2600 	/*
2601 	 * The refill handler is generated per-CPU, multi-node systems
2602 	 * may have local storage for it. The other handlers are only
2603 	 * needed once.
2604 	 */
2605 	static int run_once = 0;
2606 
2607 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_rixi)
2608 		panic("Kernels supporting XPA currently require CPUs with RIXI");
2609 
2610 	output_pgtable_bits_defines();
2611 	check_pabits();
2612 
2613 #ifdef CONFIG_64BIT
2614 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2615 #endif
2616 
2617 	if (cpu_has_3kex) {
2618 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2619 		if (!run_once) {
2620 			build_setup_pgd();
2621 			build_r3000_tlb_refill_handler();
2622 			build_r3000_tlb_load_handler();
2623 			build_r3000_tlb_store_handler();
2624 			build_r3000_tlb_modify_handler();
2625 			flush_tlb_handlers();
2626 			run_once++;
2627 		}
2628 #else
2629 		panic("No R3000 TLB refill handler");
2630 #endif
2631 		return;
2632 	}
2633 
2634 	if (cpu_has_ldpte)
2635 		setup_pw();
2636 
2637 	if (!run_once) {
2638 		scratch_reg = allocate_kscratch();
2639 		build_setup_pgd();
2640 		build_r4000_tlb_load_handler();
2641 		build_r4000_tlb_store_handler();
2642 		build_r4000_tlb_modify_handler();
2643 		if (cpu_has_ldpte)
2644 			build_loongson3_tlb_refill_handler();
2645 		else
2646 			build_r4000_tlb_refill_handler();
2647 		flush_tlb_handlers();
2648 		run_once++;
2649 	}
2650 	if (cpu_has_xpa)
2651 		config_xpa_params();
2652 	if (cpu_has_htw)
2653 		config_htw_params();
2654 }
2655